Neon: Negative Extrapolation From Self-Training Improves Image Generation
Sina Alemohammad, Zhangyang Wang, Richard Baraniuk
Instead of simply fine-tuning a generative model on its own synthetic outputs, briefly fine-tune it to find the direction of model collapse, then apply the reverse of that update to the original model for a major performance boost.
Abstract
Scaling generative AI models is bottlenecked by the scarcity of high-quality training data. The ease of synthesizing from a generative model suggests using (unverified) synthetic data to augment a limited corpus of real data for the purpose of fine-tuning in the hope of improving performance. Unfortunately, however, the resulting positive feedback loop leads to model autophagy disorder (MAD, aka model collapse) that results in a rapid degradation in sample quality and/or diversity. In this paper, we introduce Neon (for Negative Extrapolation frOm self-traiNing), a new learning method that turns the degradation from self-training into a powerful signal for self-improvement. Given a base model, Neon first fine-tunes it on its own self-synthesized data but then, counterintuitively, reverses its gradient updates to extrapolate away from the degraded weights. We prove that Neon works because typical inference samplers that favor high-probability regions create a predictable anti-alignment between the synthetic and real data population gradients, which negative extrapolation corrects to better align the model with the true data distribution. Neon is remarkably easy to implement via a simple post-hoc merge that requires no new real data, works effectively with as few as 1k synthetic samples, and typically uses less than 1\% additional training compute. We demonstrate Neon’s universality across a range of architectures (diffusion, flow matching, autoregressive, and inductive moment matching models) and datasets (ImageNet, CIFAR-10, and FFHQ). In particular, on ImageNet 256x256, Neon elevates the xAR-L model to a new state-of-the-art FID of 1.02 with only 0.36\% additional training compute.
Neon inverts model degradation from self-training by extrapolating away from it, improving generative models with minimal compute.
- Introduces Neon, a post-hoc method that reverses gradient updates from self-training to counter model autophagy
- Proves that mode-seeking samplers create predictable anti-alignment between synthetic and real data gradients
- Shows negative extrapolation corrects sampler bias and enhances recall and generation fidelity
- Achieves state-of-the-art FID of 1.02 on ImageNet 256x256 with only 0.36% additional compute
- Gradient extrapolation
- Self-training
- Diffusion models
- Flow matching
- ImageNet 256x256
- CIFAR-10
- FFHQ
Authors did not state explicit limitations.
Identify diversity-promoting samplers enabling positive alignment between synthetic and real data for direct self-improvement
from the paperActively synthesize optimal bad datasets eliciting degradation direction maximizing corrective signal
from the paper
Author keywords
- Generative Models
- Self-Improvement
- Weight Merging
- Image Generation
Related orals
Universal Inverse Distillation for Matching Models with Real-Data Supervision (No GANs)
RealUID provides universal distillation for matching models without GANs, incorporating real data into one-step generator training.
GLASS Flows: Efficient Inference for Reward Alignment of Flow and Diffusion Models
GLASS Flows samples Markov transitions via inner flow matching models to improve inference-time reward alignment in flow and diffusion models.
Generative Human Geometry Distribution
Introduces distribution-over-distribution model combining geometry distributions with two-stage flow matching for human 3D generation.
Cross-Domain Lossy Compression via Rate- and Classification-Constrained Optimal Transport
Cross-domain lossy compression unifies rate and classification constraints via optimal transport framework.
NextStep-1: Toward Autoregressive Image Generation with Continuous Tokens at Scale
NextStep-1 achieves state-of-the-art autoregressive text-to-image generation by modeling continuous image tokens with lightweight flow matching instead of diffusion.